13

As a compromise to this question, can we claim that the following symmetric complex matrix \begin{align*} S = \begin{pmatrix} 0 & 1 + 2i & 1 + 3i & \cdots & 1 + ni \\ 1 + 2i & 0 & 2 + 3i & \cdots & 2 + ni \\ 1 + 3i & 2 + 3i & 0 & \cdots & 3 + n i \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ 1 + ni & 2 + ni & 3 + ni & \cdots & 0 \end{pmatrix} \end{align*} is non-singular? To clarify, the $(k, l)$ entry of $S$ is \begin{align*} s_{kl} = \begin{cases} k + il & l > k, \\ 0 & l = k, \\ s_{lk} & k > l. \end{cases} \end{align*}

Through some row/column reduction, I have shown that \begin{align*} \det(S) = \begin{vmatrix} 0 & 1 + 2i & i & \cdots & i & i \\ 1 + 2i & -2(1 + 2i) & 2 + 2i & & \\ i & 2 + 2i & -2(2 + 3i) & \ddots & \\ \vdots & & \ddots & \ddots & \ddots & \\ i & & & (n - 2) + (n - 2)i & -2[(n - 2) + (n - 1)i] & (n - 1) + (n - 1)i \\ i & & & & (n - 1) + (n - 1)i & -2[(n - 1) + ni] \end{vmatrix}. \tag{$*$} \end{align*} Then I got stuck here. Is there any clever way to get around this? Or demonstrating the invertibility through other angles (say, linear independence or eigenvalues)?


Result updates:

Following @user8675309's tip, we can show by Levy-Desplanques theorem that the lower $(n - 1) \times (n - 1)$ submatrix of $(*)$ is non-singular, as it is row diagonally dominant. Thus $\mathrm{rank}(S) \geq n - 1$. But the passage from $n - 1$ to $n$ seems quite difficult.

Zhanxiong
  • 14,040
  • I only had a very brief scan - but maybe an inductive argument could help you here? – Riemann'sPointyNose Dec 21 '20 at 17:41
  • @Riemann'sPointyNose Thanks. I guess the passage from the inductive hypothesis to the conclusion is still tricky? One obvious route is through the Schur formula $\det(S) = \det(S_{11})\det(-S_{21}S_{11}^{-1}S_{12})$, but the difficulty still persists to show the second determinant is non-zero. – Zhanxiong Dec 21 '20 at 17:47
  • Nevermind ignore the commend I deleted, I misread the matrix values – Riemann'sPointyNose Dec 21 '20 at 18:14
  • I'm not sure I understand the pattern of the matrix S actually, could you clarify a bit more? – Riemann'sPointyNose Dec 21 '20 at 18:16
  • @Riemann'sPointyNose See my edited question. – Zhanxiong Dec 21 '20 at 18:27
  • 3
    Instead of using $R^TSR$ it may be easier to use $RSR^T$. I suspect what you want here is to know the rank of this matrix, as that specifies the congruence information for a complex symmetric (non-Hermitian) bilinear form. – user8675309 Dec 21 '20 at 19:44
  • @user8675309 Exactly. So I am guessing the rank is $n$, which is equivalent to the matrix is non-singular. – Zhanxiong Dec 21 '20 at 19:45
  • if you look at the leading $n-1\times n-1$ principal submatrix of $RSR^T$, it seems like you should be able to apply Gerschgorin discs to show that is nonsingular, implying that the rank of your matrix is at least $n-1$ – user8675309 Dec 21 '20 at 20:20
  • Does the submatrix you mentioned correspond to the lower $(n - 1) \times (n - 1)$ matrix in my question? If so, the first row and the last row seem can't rule out $0$ eigenvalue. – Zhanxiong Dec 21 '20 at 20:27
  • Can't we use Schur Complement ? Shall I give it a try ? – Darshit Sharma Jun 12 '23 at 17:11
  • Is there somewhere I can buy this "famous" book ? – dezdichado Jun 12 '23 at 17:46
  • 1
    @dezdichado https://list-chaoshi.jd.com/Search?keyword=%E7%BA%BF%E6%80%A7%E4%BB%A3%E6%95%B0%E6%9D%8E%E7%82%AF%E7%94%9F&enc=utf-8&spm=2.1.12. Thank you for your interest. A reminder is that it is written in Chinese :). – Zhanxiong Jun 12 '23 at 17:51
  • so there is no hope for an English translation ? Anyway, the remark about $\text{rank} A\geq n-1$ reminds me of this famous post https://math.stackexchange.com/questions/3906835/whats-the-sign-of-det-left-sqrti2j2-right-1-le-i-j-le-n/3954342#3954342. The solution might inspire something for this problem. – dezdichado Jun 12 '23 at 18:41
  • @dezdichado Thank you, I will take a look and try to find some connections. – Zhanxiong Jun 12 '23 at 20:33
  • @RodrigodeAzevedo Thank you, I will take a look and try to find some connections. – Zhanxiong Jun 12 '23 at 20:35

2 Answers2

6

A proof

If $n=1$, $S=(0)$ is singular.
If $n=2$, $S=\pmatrix{0&1{+}2i\\1{+}2i&0}$ is nonsingular.
If $n=3$, $\det(S)=-50-10i\not=0$.

Assume $n\ge4$.

Modify $S$ as follows. Each modification will not change whether $S$ is singular or not. Each modification will keep $S$ symmetric.

  1. For each $k$ from $n-1$ down to $1$ inclusive subtract row $k$ from row $k{+}1$ and then subtract column $k$ from col $k{+}1$.

    What is great is $S$ is now tridiagonal except on the first row and first column. $$\pmatrix{ 0 & 1{+}2i & i & i&\cdots & i & i \\ 1{+}2i & {-}2{-}4i & 2{+}2i & 0&\cdots & 0 & 0 \\ i & 2{+}2i & {-}4{-}6i & 3{+}3i&\cdots & 0 & 0 \\ \vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\ i & 0 & 0 & 0&\cdots & -2(n{-}2){-}2(n{-}1)i & (n{-}1){+}(n{-}1)i \\ i & 0 & 0 & 0 &\cdots & (n{-}1){+}(n{-}1)i & {-}2(n{-}1){-}2ni\\ }$$

  2. Subtract $\frac{-1{+}i}2\times$ the sum of all other columns from column $1$. Then subtract $\frac{-1{+}i}2\times$ the sum of all other rows from row $1$. $$\pmatrix{ 1{+}ni & i & 0 & 0&\cdots & 0 & -n \\ i & {-}2{-}4i & 2{+}2i & 0&\cdots & 0 & 0 \\ 0 & 2{+}2i & {-}4{-}6i & 3{+}3i&\cdots & 0 & 0 \\ \vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\ 0 & 0 & 0 & 0&\cdots & -2(n{-}2){-}2(n{-}1)i & (n{-}1){+}(n{-}1)i \\ {-}n & 0 & 0 & 0 &\cdots & (n{-}1){+}(n{-}1)i & {-}2(n{-}1){-}2ni\\ }$$ Now all entries of $S$ on row $1$ or column $1$ are $0$ except $5$ of them.

  3. Subtract $\frac{i}{1{+}ni}\times$ row $1$ from row $2$. Then subtract $\frac{i}{1{+}ni}\times$ column $1$ from column $2$. $$\pmatrix{ 1{+}ni & 0 & 0 & 0&\cdots & 0 & -n \\ 0 & {-}2{-}4i{+}\frac1{1{+}ni} & 2{+}2i & 0&\cdots & 0 &\frac{ni}{1{+}ni}\\ 0 & 2{+}2i & {-}4{-}6i & 3{+}3i&\cdots & 0 & 0 \\ \vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\ 0 & 0 & 0 & 0&\cdots & {-}2(n{-}2){-}2(n{-}1)i & (n{-}1){+}(n{-}1)i\\ {-}n & \frac{ni}{1{+}ni}& 0 & 0 &\cdots & (n{-}1){+}(n{-}1)i & {-}2(n{-}1){-}2ni\\ }$$ $2$ entries among those $5$ entries are now $0$ too.

    Let us show that $S$ is diagonally-dominant, which must be nonsingular.

    • Row $1$: $\ |{-}n|<|1+ni|$.
    • Row $2$: $\ |2+2i| +\left|\frac{ni}{1{+}ni}\right|$ $<\sqrt8 +1 <\sqrt{20}-\frac14$ $<|{-}2{-}4i| - \left|\frac1{1{+}ni}\right|$ $\le\left|{-}2{-}4i+\frac1{1{+}ni}\right|.$
    • Row $r=3,4,\cdots, n-1$: $\ $the nonzero elements are $r+ri$, ${-}2r-2(r+1)i$ and $(r+1)+(r+1)i$, where $-2r-2(r+1)i$ is on the diagonal. $$\begin{aligned} &\quad|r+ri| + |(r+1)+(r+1)i|=(2r+1)\sqrt2 \\ &<\sqrt{8r^2+8r+4}=|{-}2r-2(r+1)i| \end{aligned}$$
    • Row $n$: $$\begin{aligned} &\quad|{-}n|+ \left|\frac{ni}{1{+}ni}\right|+|(n{-}1){+}(n{-}1)i| \\ &< (n+1)+(n{-}1)\sqrt2\\ &= \sqrt{3n^2-2n+3+2\sqrt2(n^2-1)}\\ &< \sqrt{3n^2-2n+3+3(n^2-1)}\\ &= \sqrt{8n^2-8n+4-2(n-1)(n-2)}\\ &< \sqrt{8n^2-8n+4}\\ &= |{-}2(n{-}1){-}2ni| \end{aligned}$$

Hence the original $S$ is nonsingular too.

How did I find my proof?

The big hint is the problem is an exercise in an introductory textbook to linear algebra. However difficult it is, there should be an elementary solution that is not too fancy. So I set out to diagonalize $S$ or as much as possible. If diagonalization was too difficult, diagonally-dominant matrices should be appealing.

Step 1 was easy.

After step 1, almost entries on row $1$ are equal. So any equal quantities are of interest, which is true in more much generality. On the other hand, I had been checking whether other rows are diagonally dominant or not. Computing the sum of numbers at my hand is a secondary nature to me, even when I did not know it should be helpful beforehand. Viola, the sums of elements in the same row are equal for most rows. $$i+(2+2i)+(-4-6i)+(3+3i)=i+(3+3i)+(-6-8i)+(4+4i)=\cdots$$ The ideas connected. These equal sums can be used to change the entries on row $1$ to $0$. Thus step 2 came to light.

Step 3 was harder. I couldn't see how one could make row $1$, $2$, $n$ diagonally dominant simultaneously. If I eliminated $S[1][2]$, there would be some nonzero popping up elsewhere not on diagonal, which I thought was unlikely to be progress. I decided to try the simple step $3$ after I had succeeded in a messier approach found with some exhaustive search with my program in Python. To my delight, it worked obviously!

Apass.Jack
  • 13,396
  • 1
  • 20
  • 33
  • Is there an "actual" reason why $s$ is nonsingular (besides the general argument that almost all square matrices are nonsingular)? – Apass.Jack Jun 13 '23 at 18:32
  • I have carefully verified every step of your amazing answer! Except there is an immaterial typo in "Row 2" modulus comparison (the term $\left|\frac{1}{1 + ni}\right|$ should be $\left|\frac{ni}{1 + ni}\right|$) -- but row 2 is nevertheless still row diagonally-dominant, everything is just so amazing! I truly appreciate your help! – Zhanxiong Jun 14 '23 at 01:51
  • The transformation in Step 2 is like a miracle. Can you elaborate how did you discover it? – Zhanxiong Jun 14 '23 at 01:53
  • @Zhanxiong Thanks for checking and accepting my answer. Please see my updated answer. – Apass.Jack Jun 14 '23 at 05:47
  • 1
    Thank you for the proof outline! This is very enlightening and generalizable! – Zhanxiong Jun 14 '23 at 14:48
  • To myself, typos: "almost entries" should have been "most entries". "Voila" should have been "Voilà". – Apass.Jack Jun 15 '23 at 05:51
4

Here is a solution taken from pages 265 and 266 of a Chinese document. The solution is not mine; I just give a translation here.

There are some mistakes in the original solution (in Chinese); I corrected them.


Suppose that $n > 1$.

Adding $-1$ times row $n-1$ to row $n$, adding $-1$ times row $n-2$ to row $n-1$, ..., adding $-1$ times row $2$ to row $1$, we yield $$ S_2 = \begin{bmatrix} 0 & 1+2 \mathrm{i} & 1+3 \mathrm{i} & \cdots & \cdots & 1+(n-1) \mathrm{i} & 1+n \mathrm{i} \\ 1+2 \mathrm{i} & -(1+2 \mathrm{i}) & 1 & \cdots & \cdots & 1 & 1 \\ \mathrm{i} & 2+3 \mathrm{i} & -(2+3 \mathrm{i}) & \ddots & & \vdots & \vdots \\ \vdots & \vdots & \ddots & \ddots & \ddots & \vdots & \vdots \\ \vdots & \vdots & & \ddots & \ddots & 1 & 1 \\ \mathrm{i} & \mathrm{i} & \cdots & \cdots & (n-2)+(n-1) \mathrm{i} & -((n-2)+(n-1) \mathrm{i}) & 1 \\ \mathrm{i} & \mathrm{i} & \cdots & \cdots & \mathrm{i} & (n-1)+n \mathrm{i} & -((n-1)+n \mathrm{i}) \\ \end{bmatrix}. $$ Note that $\det {(S_2)} = \det {(S)}$.

Adding $-1$ times column $n-1$ to column $n$, adding $-1$ times column $n-2$ to column $n-1$, ..., adding $-1$ times column $2$ to column $1$, we yield $$ S_3 = \begin{bmatrix} 0 & -\alpha_{1} & \mathrm{i} & \cdots & \cdots & \mathrm{i} & \mathrm{i} \\ -\alpha_{1} & 2 \alpha_{1} & 2 \alpha & 0 & \cdots & 0 & 0 \\ \mathrm{i} & 2 \alpha & 2 \alpha_{2} & \ddots & \ddots & \vdots & \vdots \\ \vdots & 0 & \ddots & \ddots & \ddots & 0 & 0 \\ \vdots & \vdots & \ddots & \ddots & \ddots & (n-2) \alpha & 0 \\ \mathrm{i} & 0 & \cdots & 0 & (n-2) \alpha & 2 \alpha_{n-2} & (n-1) \alpha \\ \mathrm{i} & 0 & \cdots & 0 & 0 & (n-1) \alpha & 2 \alpha_{n-1} \end{bmatrix}, $$ in which $\alpha = 1 + \mathrm{i}$, $\alpha_k = -(k + (k + 1)\mathrm{i})$ for $k = 1$, $2$, $\dots$, $n-1$. Note that $\det {(S_3)} = \det {(S_2)} = \det {(S)}$.

Multiplying row $1$ by $\alpha$, multiplying column $1$ by $\alpha$, we yield $$ S_4 = \begin{bmatrix} 0 & -\alpha_{1} \alpha & \mathrm{i} \alpha & \cdots & \cdots & \mathrm{i} \alpha & \mathrm{i} \alpha \\ -\alpha_{1} \alpha & 2 \alpha_{1} & 2 \alpha & 0 & \cdots & 0 & 0 \\ \mathrm{i} \alpha & 2 \alpha & 2 \alpha_{2} & \ddots & \ddots & \vdots & \vdots \\ \vdots & 0 & \ddots & \ddots & \ddots & 0 & 0 \\ \vdots & \vdots & \ddots & \ddots & \ddots & (n-2) \alpha & 0 \\ \mathrm{i} \alpha & 0 & \cdots & 0 & (n-2) \alpha & 2 \alpha_{n-2} & (n-1) \alpha \\ \mathrm{i} \alpha & 0 & \cdots & 0 & 0 & (n-1) \alpha & 2 \alpha_{n-1} \end{bmatrix}. $$ Note that $\det {(S_4)} = \frac{1}{\alpha^2} \det {(S_3)} = \frac{1}{\alpha^2} \det {(S)}$.

Add all rows except the first to row $1$. Add all columns except the first to column $1$. We yield $$ S_5 = \begin{bmatrix} 2(-n+\mathrm{i}) & -1+\mathrm{i} & 0 & \cdots & \cdots & 0 & -n \alpha \\ -1+\mathrm{i} & 2 \alpha_{1} & 2 \alpha & 0 & \cdots & 0 & 0 \\ 0 & 2 \alpha & 2 \alpha_{2} & \ddots & \ddots & \vdots & \vdots \\ \vdots & 0 & \ddots & \ddots & \ddots & 0 & 0 \\ \vdots & \vdots & \ddots & \ddots & \ddots & (n-2) \alpha & 0 \\ 0 & 0 & \cdots & 0 & (n-2) \alpha & 2 \alpha_{n-2} & (n-1) \alpha \\ -n \alpha & 0 & \cdots & 0 & 0 & (n-1) \alpha & 2 \alpha_{n-1} \end{bmatrix}. $$ Note that $\det {(S_5)} = \det {(S_4)} = \frac{1}{\alpha^2} \det {(S)}$.

We will show that the determinant of $S_5$ is nonzero.

Since $$ \begin{aligned} & |2(-n + \mathrm{i})|^2 - (|{-1 + \mathrm{i}}| + |{-n\alpha}|)^2 \\ = {} & 4(n^2 + 1) - (\sqrt{2} + \sqrt{2}n)^2 \\ = {} & 2(n - 1)^2 > 0, \end{aligned} $$ we have $$ |2(-n + \mathrm{i})| > |{-1 + \mathrm{i}}| + |{-n\alpha}|. $$

Since $$ \begin{aligned} & |2\alpha_1|^2 - (|{-1 + \mathrm{i}}| + |{2\alpha}|)^2 \\ = {} & 20 - 18 > 0, \end{aligned} $$ we have $$ |2\alpha_1| > |{-1 + \mathrm{i}}| + |{2\alpha}|. $$

Since $$ \begin{aligned} & |2\alpha_k|^2 - (|k\alpha| + |{(k+1)\alpha}|)^2 \\ = {} & 4(k^2 + (k+1)^2) - (\sqrt{2}k + \sqrt{2}(k+1))^2 \\ = {} & 2 > 0, \end{aligned} $$ we have $$ |2\alpha_k| > |{k\alpha}| + |{(k+1)\alpha}| $$ for $k = 2$, $3$, $\dots$, $n-1$.

According to Levy-Desplanques theorem, the determinant of $S_5$ is nonzero. Hence that of $S$ is nonzero.

Juliamisto
  • 1,300
  • 1
    It looks my solution is almost the same as this one. The major difference is that I spent many hours to craft my solution, writing hundreds of line of code to explore what might work. – Apass.Jack Jun 13 '23 at 08:11
  • 1
    @Apass.Jack I do not know how the solution I present here was firstly discovered. I can justify every fact used in the argument (I checked the case in which $n = 6$ just now and I fixed some mistakes); however, I cannot think of the method myself. – Juliamisto Jun 13 '23 at 08:36
  • I thank to your answer as well (+1), which is similar to @Apass.Jack's. I wish a could have awarded you some bounty too but unfortunately the minimal bounty that I can start now is 400... – Zhanxiong Jun 14 '23 at 01:59
  • @Zhanxiong It does not matter; I am for neither fame nor fortune. I feel good if the solution here helps. – Juliamisto Jun 14 '23 at 03:08